[PATCH] BUG on fsync/fdatasync with Ext3 data=journal

[Seiji Kihara]

Hello,

We found that fsync and fdatasync syscalls sometimes don't sync
data in an ext3 file system under the following conditions.

1. Kernel version is 2.6.6 or later (including 2.6.8.1 and 2.6.9-rc2).
2. Ext3's journalling mode is "data=journal".
3. Create a file (whose size is 1Mbytes) and execute umount/mount.
4. lseek to a random position within the file, write 8192 bytes
   data, and fsync or fdatasync.

We presume the data was not written to the corresponding disk
before returning from fsync or fdatasync syscall on the evidence
as follows:

1. The response time of fsync() and fdatasync() was extremely
   short.

   We use the "diskio" tool, which is downloadable from OSDL page
   (http://developer.osdl.jp/projects/doubt/).  The program showed
   that the response time was under 10 microseconds.  This time
   cannot be achieved with data transfer on IDE and PCI bus!

2. The IDE writing routine ide_start_dma() was not called under
   DMA enabled.

   We inserted the print messages in the sys_write(), sys_fsync()
   and ide_start_dma() by the attached patch.  Sometimes the
   "ide_start_dma: ..." message was not shown between "write: in
   ..." and "fsync: out ...".

The problem was occurred since 2.6.5-bk1, which includes the patch
"[PATCH] ext3 fsync() and fdatasync() speedup".  We found that the
problem was solved by deleting the part of the patch which
modifies ext3_sync_file().  Maybe, i_state is not correctly set to
I_DIRTY when the related page cache is dirty (is it true?)

Attached file is tarball (tar + bzip2) which contains following
files.  The patches are for 2.6.8.1-kernel (applicable to
2.6.9-rc2), and the results were also produced with
2.6.8.1-kernel.

- fsync.c.patch: patch for solving the problem
- kernel.printk.patch: patch for showing the problem by printks
- kernel.printk.log: printks from the kernel with the printk patch
- results/*: output from "diskio" program
  (result for fsync, and O_SYNC for comparison)
- kernel-2.6.8.1.config: .config file which we used to build kernel

We use the following hardware to make the data:

  CPU: Intel Pentium 4 3GHz Hyper Threading, 2nd cache 512KB
  RAM: 1GBytes
  IDE controller: on-board ICH5 (ATA100)
  HDD: E-IDE (7200rpm, 8MB cache, ATA133) x 2
       (one for system and the other for writing test data)
  (The problem was also reproduced with a SCSI system)

Regards,

Seiji

-- 
Seiji Kihara
Open Source Software Computing Project,
NTT Cyber Space Laboratories, Yokosuka, JAPAN

-------------- next part --------------
A non-text attachment was scrubbed...
Name: patches+results.tar.bz2
Type: application/octet-stream
Size: 80725 bytes
Desc: not available
URL: <http://listman.redhat.com/archives/ext3-users/attachments/20040916/869218f6/attachment.obj>